What if spawned sub-agents started in Plan Mode? They propose a plan, the orchestrator reviews it, and only then do they execute. A simple “plan → review → execute” loop could reduce agent drift, duplicated work, and confidently solving the wrong problem.